[Pw_forum] multi core Vs Single core

[Axel Kohlmeyer]

On Fri, 8 Feb 2008, ambavale sagar wrote:

 
sagar,

AS> Is multi (quad) core processor more efficient than single core for our

please clarify what you mean by "efficient".
more throughput for the same money, faster execution,
better utilization of the cpu, less room needed for 
a given computational capability?

AS> abinitio calculations? How efficiently multi-threading is done while
AS> working with serial version of code on multicore machine?

it is a common misconceptions that you need to use multi-threading
to take advantage of multicore cpus. the vendors of those cpus
are somewhat promoting it, because it is simpler to add multi-threading
to a code than to parallelize it completely, and also because
their new cpus don't have any advantage over the older ones without
paralleization. fortunately, the methods used in quantum espresso 
are well parallelizable and the code offers multiple levels of MPI
parallelism which i found to be almost always more efficient than
using OpenMP (and thus multi-threading). even more so, QE does not
support OpenMP (only indirectly via threaded BLAS/LAPACK libraries).

depending on your hardware, problem set size and compiler/library
combination, you can get a speedup of roughly 2.5x to 3.5x from
a quad core cpu over a corresponding single core. when using multiple
nodes via a network, however, the speed and the latency of the 
network (and the options of parallization, i.e. if you can parallelize
over k-points or not) start to matter and can make it all very 
confusing. 

in the end, apart from a few general observations, there is no other
way to get a definite answer than running tests with representative
input files.

cheers,
   axel. 

p.s.: this brings up the question again, that we should compile a
list of such typical inputs and run them on mutually available 
machines so people can see what to expect....

AS> 
AS> Thanx.
AS> 
AS>  
AS> 
AS> Sagar K. Ambavale
AS> 
AS> PhD student,
AS> 
AS> The M.S. University of Baroda,
AS> 
AS> India
AS> 
AS> 
AS> ________________________________________________________________________________
AS> 5, 50, 500, 5000 - Store N number of mails in your inbox. Click here.
AS> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.